Artificial intelligence for diagnostic applications in rare hematologic malignancies: A systematic review Free

Rare hematologic malignancies (collectively accounting for ~20% of cancers) pose diagnostic challenges due to heterogeneous presentations and limited expert availability. Artificial intelligence (AI), especially machine learning (ML) and deep learning, has emerged as a promising tool to improve the diagnosis and classification of these diseases. We conducted a systematic review of human studies on AI-based diagnostic tools for rare hematological malignancies (e.g. myelodysplastic syndromes, acute leukemias, lymphomas, multiple myeloma), focusing on those with retrospective clinical validation.

A comprehensive literature search (up to 2024) identified original studies applying AI/ML algorithms to the diagnosis of rare blood cancers. Inclusion criteria required human subject data and reported performance metrics (sensitivity, specificity, accuracy or AUC) from retrospective validation; many studies also compared AI models to standard diagnostic methods or expert clinicians. We found numerous AI-driven diagnostic applications across various rare hematologic malignancies. Morphologic analysis: Deep learning models applied to blood/bone marrow smears can identify malignant cells and dysplasia with high accuracy. For example, a convolutional neural network model distinguished acute promyelocytic leukemia (APL) from other acute myeloid leukemia (AML) subtypes and healthy marrow with AUC ~0.86–0.96. In myelodysplastic syndromes (MDS), image-based classifiers detected dysplastic cells in multiple lineages with AUC ≥0.94 and could differentiate MDS from look-alike conditions (aplastic anemia or AML) with ~92% accuracy (external validation AUC ~0.94). Flow cytometry and multi-omics: ML applied to multiparametric flow cytometry data has outperformed conventional scoring systems (e.g. Ogata score) for MDS diagnosis. An elastic net model using flow cytometry features achieved 93.5% AUC (92% sensitivity, 93% specificity) in an external test cohort. Similarly, a Random Forest algorithm on flow cytometry immunophenotyping data improved diagnostic accuracy for MDS (97% sensitivity and 95% specificity in validation) compared to expert-based flow scores. Other modalities: In multiple myeloma, integrated ML models using routine blood test panels distinguished precursor monoclonal gammopathy (MGUS) from active myeloma with ~93% accuracy (AUC ~0.97), while deep learning on bone marrow aspirate images plus clinical data achieved >90% diagnostic accuracy. For lymphomas, early machine learning algorithms have been explored to classify lymphoma subtypes and predict outcomes using histopathology images and genomic data, with some models reaching expert-level diagnostic performance. Across studies, many AI models matched or exceeded the accuracy of experienced hematologists in detecting malignancies. Notably, the “MOSAIC” AI framework demonstrated that multimodal learning (combining clinical and genomic features) can better capture disease heterogeneity in a rare cancer (MDS), yielding superior risk stratification compared to traditional prognostic scoring (C-index 0.77 internal, 0.74 external). AI-based diagnostic tools show promise in rare hematological malignancies, demonstrating high accuracy in identifying diseases like MDS, AML (including APL), lymphoma, and multiple myeloma. These models can integrate complex data (morphology, flow cytometry, genomics) to detect subtle patterns often missed by conventional methods, thereby facilitating earlier and more precise diagnoses. In retrospective evaluations, many ML algorithms have complemented or even outperformed standard diagnostic approaches and expert assessments. Incorporating such tools into clinical workflows could expedite diagnosis, improve risk stratification, and guide treatment decisions for rare blood cancers. However, most studies to date are retrospective; prospective clinical validation and regulatory-approved trials are needed to establish their real-world impact and safety. Challenges remain in ensuring generalizability (through external validation across diverse populations), integration with existing lab systems, and addressing ethical/regulatory issues for AI in diagnostics. In summary, AI-driven diagnostic frameworks for rare hematologic malignancies have achieved impressive performance in preliminary studies, and with further validation they hold potential to become reliable assistive tools in hematology, ultimately improving diagnostic accuracy and patient outcomes.